Grok Bot x Cursor: The Harness Era

You talk to @Bot. It starts the Cloud Agent. Cursor does the lift.

---

My POV is this.

Cursor already had remarkable Cloud Agent infra. Most of us just would not live in that tab all day.

Grok Bot is how you actually start it.

You tell the Bot the job. It launches the Cloud Agent. You review the PR.

Felt like ~1/20th the setup cost. Massive unlock.

The lane:

Somebody talks to Grok Bot. The Bot starts a Cursor Cloud Agent.

Same move. Different jobs on top of it.

The Bot should not do the coding.

Cursor plan pays the repo work. Grok Bot weekly bucket stays intact. Different meters.

Bot architects. Cursor does the heavy lift.

Every Cloud Agent runs /poteto-mode.

---

The move

Old loop: open Cursor, pick the repo, write the prompt, babysit the thread.

New loop: tell the Bot to launch a Cloud Agent. It shows up at cursor.com/agents. The Bot stays the manager. Your Cursor plan pays for the repo work. You apply judgment on the PR.

• YOU talk to one Bot

• the Bot starts / sizes / watches Cloud Agents

• those agents run /poteto-mode

• Cursor does the heavy lift

• you + your Bot review

---

1. Kun Chen — the factory

@kunchenguid owns a pile of open source (24k+ stars) and was drowning in issues and PRs.

He did not use Grok Bot to write the repos. He told a Bot called Firstmate to run a software factory. Firstmate talks to crewmates. Crewmates spin Cursor Cloud Agents so the coding does not fight for the shared Bot VM. Decisions and PRs come back to Kun.

His later receipt: 625 unique items triaged after they stood the automation up. 331 issues, 294 PRs.

Why it's good: Cloud Agents let the factory scale sideways. The Bot stays captain. Kun applies judgment. He said the quiet part — you could do this without Grok Bot, just not this easily and not this low-maintenance.

Paste this:

Template: github.com/kunchenguid/grok-ship

Template: github.com/kunchenguid/grok-ship

Setup thread: x.com/kunchenguid/status/2090463366762676732

---

2. Ian Nuttall — the Bot picks the model

@iannuttall asked how ppl use Grok Bot for dev, then answered himself.

You ask the Bot. It uses Cursor Cloud Agents. His Bot changed the default off Opus and started routing:

• Fable for planning

• Sol 5.6 for implementation

• Grok 4.6 for general purpose

All managed by the Bot. He is not hopping model pickers. RIP Sol in that stack.

OpenAI is winding down models in Cursor. Proposed shutoff Nov 12. So the implementer slot Ian just set up is already on a clock. Sol is a great model. I really like it. It's just not going to live in this harness much longer.

Doesn't change the move. You still tell the Bot which model does which job. You still launch Cloud Agents. You swap the implementer when the menu changes. That's the whole point of a harness.

A few days later he posted from a hammock in Tenerife. Grok Bot on support. Cursor background agents on bugs. Still the same primitive. The model names on the ticket are the part that moves.

Why it's good: this is the 2+2=5 I keep talking about. The harness that can route 2 or 3 models on the same job is the alpha. Cursor is that harness. The Bot is how you actually start it from your phone. Sol leaving the menu is a signal, not a reason to go back to living in the agent tab.

Paste this:

Ian wrote Sol in the implementer slot. If OpenAI models are already gone on your account, put Grok 4.6 Extra High there instead.

---

3. Nick + Brandon — managers stay managers

This is the usage hack that made juices flow for a lot of ppl.

@BStarr119 was burning the whole Grok Bot weekly bucket and none of his Cursor Ultra. He saved a skill so his Bots only manage. Heavy work launches as long-running Grok 4.6 Cloud Agents.

Result he posted: jobs that ate 5% Bot usage dropped under 1%. Same or higher throughput. Ultra pays.

@thedogfather wrote the same skill down as Heavy Lift Cloud Agents. Bots do not code, browse, or chew the VM. They scope, launch, watch, revise, deliver.

Why it's good: two meters. Grok Bot weekly usage is one. Cursor plan usage on Cursor models is the other, and that pool is generous. If the Bot codes on its own computer, you torch the wrong bill. If the Bot starts Cloud Agents, Cursor does the lift and the Bot stays cheap.

Paste this:

Guide: tlgenapp.com/gbot-skills

Guide: tlgenapp.com/gbot-skills

If a Bot ignores it, append use heavy lift to the prompt.

---

4. Kevin Parko — the eng Bots do not write the repo

@n2parko is product at SpaceXAI. He used Grok Bot internally to help build Cursor.

The line that matters: Grok Bot showed up as a double-digit share of internal merged PRs.

His eng manager Bot does not code. It breaks work down and delegates. The IC eng Bots spin Cursor Cloud Agents and check the output against the goal.

Why it's good: this is Heavy Lift with names on the Bots. Separate memory. The manager stays the manager. Cloud Agents are the ICs.

Paste this:

Longer PM setup: x.ai/bot/guides/grok-bot-for-pms

Longer PM setup: x.ai/bot/guides/grok-bot-for-pms

---

5. Krista Letz — spin more than one

@kristaletz wrote the Enterprise GTM piece. Different job than shipping PRs, same primitive.

Tell your Bot to spin up multiple Cloud Agents for parallel tasks, and make a skill for it. She also has a "Cursor 10x engineer" Bot connected to the codebase so she can answer a customer live without pinging an engineer.

Why it's good: parallel Cloud Agents are how the Bot stops feeling single-threaded. Ian asked that question out loud. Krista just runs it.

Paste this:

---

The honest limit

The unlock is real. So is the bill if you run it sloppy.

@kloss_xyz found that a Bot-launched Cloud Agent can silently flip to fast after the first reply and bill as cursor-grok-4.6-high-fast. He torched 1B tokens in 7 days.

Check the model field on the follow-up before you send. Keep the models you actually want enabled in Cloud Agent settings.

He also audited routines eating the Bot meter — a 15-minute cron is 96 runs a day, inactive Bots still hitting the web, group chats where three Bots answer the same question. $1,048 is what those habits cost him. Worth the read before you scale Kun's factory.

Other small ones I keep seeing:

• Bot-started agents can hide until you turn on the source filter in the agents list (@piotrpliszko)

• Team toggle "review cloud agents" needs to stay on

• If you never send the work to Cloud Agents, the Bot will chew its own coding bill

You'll prbly do one of these once. Then put Heavy Lift on and stop lighting the wrong meter.

---

How I run it

I tell the Bot the job. I tell it to start a Cloud Agent. I set a rule with my bot to always run cursor agents in /poteto-mode.

Not boxing out the other stacks. Fable still comes out when I need the oracle.

Grok Bot x Cursor Cloud Agents is the lowest setup-cost way I have found to get real repo work moving while I am not staring at the agent tab.

Cursor had the infra. The Bot made it something you actually start.

We're in the Harness Era.

---

Try one

Pick one of the paste blocks. Or two. Or three.

Give it to your Grok Bot. Point it at a real repo.

Tell it, every time it launches Cursor, run /poteto-mode.